SpliceAI

Implementation of SpliceAI models in Pytorch with added Dropout layers

Model was trained transcripts from hg38.

For transcripts with multiple isoforms, isoform with highest number of exons was chosen for training and testing.

Setup

Create conda/mamba environment

mamba env create -f environment.yml 
mamba activate spliceAI

Download datasets for training and testing the models

Download from Google Cloud Storage (recommended)

python3 utils/download_dataset.py

You can also recreate datasets from scratch with

python3 utils/create_datasets.py \
        --fasta # reference genome - leave as "" to download
        --gff  # reference genome gff annotations - leave as "" to download
        --chrom # chromosome id map - leave as "" to download
        --flank 5000 # 5000 if you want to prepare datasets for the largest SpiceAI model
        --outdir <Path to output directory>

Train model

python3 model/train.py 
--train_data data/datasets/train.pkl 
--val_data data/datasets/val.pkl 
--train_fraction 0.02 # set to 1. to train on entire dataset
--val_fraction 0.2 # set to 1. to validate on entire dataset
--batch_size 256 
--num_epochs 200 
--learning_rate 0.001 
--flank 5000 # 5000 is for SpliceAI_10k model
--seed 1809 
--output_dir <Path to output directory to save model checkpoints and log file>

Result

SpliceAI_10k model

Loss curves

Metrics